Introduction
The Homeland Security AI Guidelines are no longer a single framework but a layered stack of overlapping policy documents issued across DHS. The 2024 roadmap set direction, the April 2024 critical infrastructure guidelines added the first security floor, and Directive 139-08 locked in internal governance in early 2025. The September 2025 compliance plan for Office of Management and Budget memo M-25-21 turned those ambitions into deadlines with teeth. According to the department, DHS now operates more than 180 AI use cases, with law enforcement and biometric matching leading the catalog. The guidelines are, in effect, the operating manual for the fastest-growing AI estate inside the federal government. For agencies, contractors, and critical infrastructure teams the practical question has shifted toward proving compliance by 2026 and 2027. This article unpacks the entire stack, explains where the guidelines actually bite, and offers a usable path to compliance.
Quick Answers on Homeland Security AI Guidelines
What are the Homeland Security AI Guidelines?
The Homeland Security AI Guidelines are the layered set of DHS policies, OMB memos, and CISA documents that govern how federal and critical infrastructure AI systems are built, acquired, assessed, and operated across the United States.
Who must comply with the Homeland Security AI Guidelines?
Every DHS component, their contractors, and operators of the sixteen critical infrastructure sectors must comply with the Homeland Security AI Guidelines when their systems make, assist, or document decisions that affect safety, rights, or security.
Key Takeaways on the Current Guidance
- The current Homeland Security AI Guidelines sit on four pillars: the DHS AI Roadmap, the critical infrastructure framework, Directive 139-08, and the OMB M-25-21 compliance plan.
- Every DHS AI system is now sorted into risk tiers, with high-impact AI triggering mandatory Impact Assessments and Chief AI Officer sign-off.
- Operators of critical infrastructure face new agentic AI and operational technology security rules from CISA that most legacy SOC playbooks do not cover.
- Civil rights, biometric, and Section 702 interactions remain the sharpest unresolved tensions and the most likely source of litigation through 2027.
Table of contents
- Introduction
- Quick Answers on Homeland Security AI Guidelines
- Key Takeaways on the Current Guidance
- Understanding Homeland Security AI Guidelines in Federal Programs
- How the DHS AI Roadmap Became the Operating Spine of Agency Policy
- The Role of OMB M-25-21 in Shaping Agency Compliance Plans
- DHS Directive 139-08 and the Chief AI Officer Mandate
- How CISA Shapes AI Security Rules for Critical Infrastructure Operators
- Agentic AI Guidance and the New Threat Surface for Operational Technology
- Impact Assessments, Risk Tiers, and the High-Impact AI Designation
- How to Implement a Compliant DHS AI Program Inside Your Organization
- Biometric Expansion at the Border and the Mobile Fortify Flashpoint
- Immigration Enforcement, ICE, and the Civil Rights Debate Around AI
- AI in Transportation Security and the TSA Pilot Portfolio
- Procurement Rules, Vendor Attestations, and the Shift to Enterprise AI Gateways
- The Signals Intelligence Boundary and FISA 702 Interactions
- Red Teaming, Evaluation, and Documentation Expected of DHS AI Owners
- The Future of Homeland Security AI Guidelines Through 2027 and Beyond
- Key Insights on the Current Framework
- Real-World Examples of DHS AI Deployments Under the Current Guidelines
- Case Studies That Reveal How the Guidelines Hold Up Under Pressure
- Risks, Ethical Tensions, and Oversight Gaps in the Current Framework
- Frequently Asked Questions on Homeland Security AI Guidelines
Understanding Homeland Security AI Guidelines in Federal Programs
Homeland Security AI Guidelines are the layered DHS policies, OMB memos, and CISA documents that direct how federal and critical infrastructure AI systems must be built, acquired, assessed, documented, and operated across the United States.
An Interactive From AIplusInfo
DHS AI Compliance Explorer
Pick a component, a system type, and a tier assumption to see which Homeland Security AI Guideline duties apply and how long the compliance work should take.
Immigration and Customs Enforcement
Investigative drafting assistant
High-impact (Tier 1)
Compliance footprint for the selected system
Required
14 to 20 weeks
Yes
Signed by Privacy Officer
High-impact enforcement AI inside ICE requires a full Impact Assessment, a civil rights review, red team coverage, and quarterly AI Oversight Board review.
Values derived from the DHS compliance plan for OMB M-25-21 and Directive 139-08.
How the DHS AI Roadmap Became the Operating Spine of Agency Policy
The DHS AI Roadmap, first published in March 2024 and refreshed through 2025, is the backbone that every other Homeland Security AI Guideline hangs on. The roadmap organizes DHS AI work into three lines of effort covering responsible use, workforce, and partnership with industry and civil society. It is not a regulation on its own, but every DHS component now writes its annual AI plan against the roadmap’s priorities. Three high-visibility pilots, including the ICE investigation drafting assistant and the FEMA hazard mitigation support tool. Trace back to the roadmap and have become references for the rest of the department. The Digital Government Hub summary gives the public-facing version in one page.
Beyond the pilots, the roadmap pushes DHS to publish an AI use case inventory that any member of the public can read. Turned out to be the first time the federal government disclosed enforcement AI at this scale. The inventory now covers more than 180 use cases across components, with the ICE and CBP catalogs growing fastest. It anchors the broader conversation about AI governance trends and regulations. Critics argue the roadmap under-specifies civil rights safeguards, and that complaint has driven the sharper language that later appeared in Directive 139-08. The practical reading for a vendor is that any new DHS proposal must map to a roadmap line of effort or it is unlikely to be funded.
The Role of OMB M-25-21 in Shaping Agency Compliance Plans
Building on the roadmap, the Office of Management and Budget memo M-25-21 reset the federal AI compliance baseline in April 2025 and forced every agency. Including DHS, to publish a formal compliance plan by late September 2025. The memo replaced the prior M-24-10 framework and sharpened three obligations: a public AI use case inventory, a documented process for identifying high-impact AI, and minimum practices for pre-deployment testing. DHS answered in a 68-page compliance plan that explicitly ties each M-25-21 obligation to an internal owner, usually the Chief AI Officer or the component Chief Information Officer. That document is still the clearest window into how DHS plans to prove it meets the memo.
M-25-21 also forced DHS to publish waiver procedures, so that a component that cannot meet a minimum practice must formally request a waiver, record the risk, and name a mitigation. The memo introduces a strict split between rights-impacting and safety-impacting AI, with different minimum practices for each. Both categories collapse into the DHS “high-impact” label used across its use case inventory. Enterprise contractors should expect those terms to appear in solicitations, statements of work, and award documents through 2026 and 2027. Analysis by Kiteworks on the DHS AI strategy describes how data governance controls have to flow all the way to the vendor layer.
One underreported effect of M-25-21 is the shift toward enterprise AI gateways, meaning a controlled interface through which all federal AI calls pass. DHS has committed to a departmental gateway that logs prompts, enforces data loss prevention, and attaches minimum practice evidence to every call. The reasoning is that policy without instrumentation does not survive a real audit, and that each component cannot be trusted to build a gateway of its own. Vendors who already run model gateways will find themselves closer to compliant than those who sell raw model APIs. A dynamic that mirrors what we have seen across broader White House AI regulation rules.
Finally, M-25-21 creates a feedback loop where agency compliance plans themselves become public documents, which other agencies read and copy. The DHS plan has influenced the Department of the Interior and the National Archives plans, which arrived weeks later and borrow the same high-impact classification language. For a vendor this means a question answered in the DHS context in April 2026 may surface again in July 2026. The same wording tends to appear in a Department of Energy solicitation. That is the practical reason OMB M-25-21 should be read as the center of gravity for every federal AI RFP for the next two cycles. The practical takeaway is that program managers who ignore this detail usually see it return as a corrective action.
DHS Directive 139-08 and the Chief AI Officer Mandate
Stepping inside DHS, Directive 139-08 took effect in January 2025 and is the single most binding internal document in the policy stack. The directive formalizes the Chief AI Officer role, assigns an AI Oversight Board, and defines prohibited uses of AI inside the department. Prohibited uses include unlawful surveillance, decisions that override a human on a protected class basis, and automated adjudication of asylum or benefits applications without human review. These are hard lines, not aspirational statements, and a program manager who crosses one is subject to administrative action under existing DHS personnel rules. That shift is quiet but material for anyone preparing a program schedule.
The directive also requires every component to maintain an AI use case register that is reviewed quarterly and reconciled with the public inventory. Closed a transparency gap flagged by the Brennan Center analysis of DHS AI transparency. The quarterly cadence is what finally made the ICE and CBP inventories visible to outside analysts in roughly real time. Civil society groups still argue the register undercounts investigative tools, and that argument continues through letters to Congress and FOIA litigation. Compliance teams should treat the register as a living document and build a change log, not a static spreadsheet.
Directive 139-08 is also the first DHS document to tie AI governance to acquisition in a way that reaches vendors directly. New contracts above a defined threshold must include an AI data governance clause, an evaluation clause, and a notification clause that triggers if the vendor materially changes a model. This mirrors internal controls private firms have adopted under their own responsible AI programs, as explored in our piece on responsible AI governance frameworks. In practice, if your product rides a weekly model refresh, you now need a workflow for notifying DHS before the refresh ships, with evidence that your evaluation suite still passes. Vendors who treat the obligation as a release artifact rather than a one-time filing come out ahead. The pattern is now consistent across components and should be assumed as the default posture.
How CISA Shapes AI Security Rules for Critical Infrastructure Operators
Turning to critical infrastructure, the Cybersecurity and Infrastructure Security Agency sits inside DHS and sets the AI security guidance that every pipeline operator. Water utility, electric utility, hospital system, and port terminal is expected to follow. The April 2024 safety and security guidelines were the first door, and the 2025 and 2026 joint guides on operational technology have become the practical reference. These documents are nonbinding in a legal sense, but they are cited by insurers, examiners, and plaintiffs, which turns them into de facto rules. Industry publications such as Industrial Cyber’s framework analysis have become the fastest way to track updates.
The CISA guidance is organized around four roles: cloud and compute providers, AI developers, critical infrastructure owners. Civil society plays a formally defined role in the four-role architecture here. Each role carries distinct obligations around secure development, model supply chain accountability, data integrity, and incident sharing. The guidance pushes operators to log every AI interaction with safety-relevant systems and to maintain an incident reporting lane that terminates in CISA. Is a new mechanism and one that most operators have not funded yet. The practical read is that if your SOC does not know which models touch your SCADA environment, you are already behind. That answer is as much about governance as it is about AI and cybersecurity.
Agentic AI Guidance and the New Threat Surface for Operational Technology
Shifting focus to the newest guidance, the 2026 agentic AI document from CISA and allied agencies addresses a threat surface that did not exist when the roadmap was written. Agentic AI systems act with partial autonomy, can call tools, can read and write to systems. Can chain actions across a session, which breaks the assumption that an AI model only outputs text. The guidance pushes operators to inventory every tool an agent can call and to require explicit allowlists for actions that touch safety or financial systems. It also sets a strong expectation that agent actions are logged in a tamper-evident store, which many legacy operational technology shops cannot produce today.
Beyond logging, the agentic AI guidance treats confused deputy and prompt injection as first-class risks and names specific mitigation patterns. These patterns include per-tool scoping, time-bounded delegation, and human-in-the-loop confirmation for any destructive action, mirroring approaches we have covered in agentic AI for enterprise security. Operators who skip these patterns and run agent frameworks with full read and write access to production systems are the headline incident waiting to happen. CISA cites real-world near-miss cases from 2025 and 2026, including an electric utility whose agent framework began writing to its historian database before being paused. Teams that build this reality into their roadmap avoid the painful retrofit later.
Operational technology environments are notoriously harder to patch than enterprise IT. This is where the guidance departs from standard IT advice and recognizes that an operator may not be able to upgrade a controller for years. The recommended compensating controls include network segmentation on a per-agent basis, strict data diodes between IT and OT, and offline model evaluation using production-representative data captures. These controls connect to broader themes in autonomous AI oversight frameworks, and they anchor the next 12 months of OT security investment plans across the energy and water sectors. That is the pattern most inspectors general reference during a lookback review. Programs that invest early in this muscle report fewer surprise findings during their next cycle.
Impact Assessments, Risk Tiers, and the High-Impact AI Designation
Looking at the way DHS actually sorts AI systems, the high-impact designation is now the single most consequential label in the guidelines. It determines what documentation, review, and sign-off a system needs. A system is tagged high-impact when its decisions materially affect rights, safety, access to benefits, or law enforcement outcomes. The inventory includes facial recognition, asylum pre-screening, and benefit adjudication tools among its top entries. Analysts at the American Immigration Council on high-impact AI have detailed how ICE classifies its tools and where the department has drawn lines critics find uncomfortably narrow. The classification choice drives whether a system needs an Impact Assessment or merely a lighter use case registration.
The Impact Assessment itself is a structured document that captures the problem, the model, the training data, the evaluation results. A civil rights review signed by the component privacy officer. Assessments must be re-run on material change, which means a weekly or monthly refresh model cannot ship under an eighteen-month-old assessment, a fact many vendors still miss. Our prior coverage of AI data readiness framework patterns is directly relevant because the assessment leans heavily on data documentation that most teams only build after they are asked. The lesson for 2026 and 2027 is to design for scrutiny from day one, not from the first audit.
DHS maintains three risk tiers, with the top tier reserved for autonomous or rights-impacting enforcement decisions. The middle tier covers decision support where a human reviewer is in the loop. The lowest tier covers productivity or research tools with no decision authority. Each tier has its own minimum practices and its own cadence of continuous monitoring, with the top tier reviewed quarterly by the AI Oversight Board. The specificity matters because a mistier can convert what should be a six-week registration into a six-month assessment, blowing a program schedule. Market analysts expect this trend to accelerate as rulemaking tightens the loop.
The last point to make here is that tiering is not static, so an initially low-tier productivity tool can graduate to high-impact when it is integrated into an enforcement workflow. DHS has already walked a FEMA assistance scoring tool and an ICE investigative drafting assistant up the tier ladder after legal review flagged downstream enforcement use. These transitions are the quiet source of most compliance pain, since the tool team usually has no new budget to run the heavier controls. Teams that build with the top tier in mind from the start carry less migration risk and tend to clear the Impact Assessment faster. The guidance is still evolving, so document any choice carefully so your next revision can trace it.
How to Implement a Compliant DHS AI Program Inside Your Organization
Stepping back from the policy stack, a compliant DHS AI program inside a federal agency or contractor organization has five concrete pillars that reviewers will probe. Those pillars are an inventory, an assessment process, an evaluation suite, a monitoring loop, and a decommissioning plan, and the guidelines ask for evidence on all five. Most organizations we have reviewed can show the first three in some form but stumble on monitoring and decommissioning. That gap is where a theoretical compliance plan becomes a demonstrable one or quietly fails during an inspector general audit.
The most practical starting move is a one-week scoping sprint that maps every AI touchpoint in the business to a tier, a system owner, and a last-reviewed date. That spreadsheet, imperfect as it is, becomes the backbone for Impact Assessments, waiver requests, and vendor flow-down requirements. Teams that start with a vendor-driven controls catalog usually invert the problem and end up spending budget on controls that do not match any real risk. Our prior coverage of a Colorado AI Act compliance guide shows how similar mapping discipline unlocks a state-level program and the same principle applies here.
Beyond mapping, the single most undervalued control is a continuous monitoring loop that compares live model behavior against the baseline captured during the Impact Assessment. The loop should include drift metrics, prompt and response sampling, and a quarterly tabletop that walks a mock enforcement decision from input to redress. These habits, grounded in principles from our AI ethics and laws coverage, are what turn a compliance binder into a working program. Inspectors general and the AI Oversight Board can tell the difference between a program that lives in SharePoint and one that lives in production.
Biometric Expansion at the Border and the Mobile Fortify Flashpoint
Shifting to the border, the biometric expansion inside DHS has turned Mobile Fortify into the sharpest public test of the Homeland Security AI Guidelines. The application is a handheld face-matching tool used by immigration officers. Bloomberg reported it reads against a database of 1.2 billion images stitched together from passport, visa, and border crossing records. Reporting from Bloomberg on the DHS face-scanning app framed the capability as the largest known operational face-matching deployment inside the federal government. Civil rights groups argue the tool is high-impact under the DHS classification scheme and should trigger the heaviest documentation and review.
Beyond Mobile Fortify, the December 2025 final rule for the Biometric Entry/Exit program made face capture a standard part of air, sea. Land border crossing, pulling millions of records into the system each week. That expansion sits alongside pilot work on AI smart glasses for field officers and renewed debate about New Orleans facial recognition debate in municipal policing. Operators and vendors should assume every border biometric project will be read as high-impact AI by the time the next inventory refreshes. The guideline-driven answer is to publish a thorough Impact Assessment up front rather than retrofit one under subpoena.
Immigration Enforcement, ICE, and the Civil Rights Debate Around AI
Turning to Immigration and Customs Enforcement, the agency’s AI estate has grown faster than any other DHS component and is where the the current guidelines face their sharpest stress test. ICE uses natural language models to draft investigation reports, risk tools to prioritize detainer leads, and vendor platforms such as Palantir for case analytics. Analysis from the American Immigration Council on invisible gatekeepers argues that too many of these tools are classified below high-impact, which lets ICE skip the deeper Impact Assessment. The classification argument is now the central legal debate inside the civil rights community. That operational reality is what separates mature programs from paper-only ones.
ICE insists its AI tools are decision support rather than decision makers. The agency assigns many of them to the middle tier. That position is contested in several active lawsuits, and the court record will likely force a tier recalibration before 2027. Policy debates around AI and policing issues echo almost identical arguments about who truly makes a decision when a human reviewer rubber-stamps an AI output. The guidance offers the human-in-the-loop test, but critics argue that test is routinely satisfied in form rather than substance. Component leaders who have lived through one Impact Assessment cycle rarely want to repeat it unprepared.
The civil rights conversation is not purely abstract because the inventory shows real tools with real caseload impact. Including an investigative drafting assistant that handled roughly 15 percent of ICE field investigation reports in late 2025. That volume means a model failure at scale can propagate into enforcement actions long before a quarterly review catches it. The current guidelines call for a complaint and redress channel for individuals affected by AI decisions, but the channel is new and uptake has been slow. Expect litigation to force the design of that channel faster than internal reform does.
AI in Transportation Security and the TSA Pilot Portfolio
Shifting to transportation, the Transportation Security Administration runs the most visible civilian AI pilots inside DHS, including automated identity verification at airport checkpoints and risk-based screening pilots. The identity verification system is the one most Americans will encounter, and it compares a traveler’s face against the real-time image captured at the podium. The pilots proceed under a formal privacy impact assessment and a public notice, and TSA has committed to permitting travelers to opt out and present a human-reviewed ID instead. That option is central to how the guidelines translate into a touchable civil liberties guarantee. The next revision cycle will probably codify what the strongest teams already practice.
Beyond identity verification, TSA has invested in explosive detection models for computed tomography scanners and in behavior detection analytics, with mixed results that have not yet settled into deployment doctrine. Audits have pushed TSA toward tighter evaluation discipline, which echoes the broader shift in the department. Analytical patterns from AI and data surveillance apply directly, since the surveillance-flavored tools are the ones under the most scrutiny. For vendors, selling into TSA now requires a bias and error-rate disclosure that most physical security firms do not routinely produce. Compliance that is only visible in a binder tends to crumble when a line officer asks a hard question.
Procurement Rules, Vendor Attestations, and the Shift to Enterprise AI Gateways
Moving to procurement, DHS has leaned hard into contract clauses as the enforcement vehicle for the guidelines because contracts bind vendors in a way policy cannot. New solicitations above a defined ceiling include a mandatory AI data governance clause, a model update notification clause. An evaluation clause that requires the vendor to re-run a standard test suite on each release. A reader can see the shape of these obligations in CBP Directive 1450-030 on AI operations, which parallels Directive 139-08 at the component level. Vendors who read these documents late tend to redline themselves out of competition.
Alongside the contract clauses, DHS is moving toward enterprise AI gateways as the preferred acquisition pattern, which changes how vendors price and package their products. In a gateway world the vendor provides a model and an evaluation harness, and the gateway owns audit logging, data loss prevention, and policy enforcement. That shifts significant engineering effort from the vendor to the department, but it also raises the bar on evaluation discipline. Teams that already treat evaluation as a product feature, as our overview of AI data readiness framework suggests, are naturally better positioned to win gateway-based awards. This is where governance and engineering must share the same calendar, not two separate ones.
Procurement is also where the guidelines bite small vendors hardest, because the attestations and evaluation suites assume engineering teams of a size most startups do not have. DHS is piloting a pre-qualification pathway that smaller vendors can use to pre-run a model evaluation and attach the evidence to a bid. The pathway is new, uneven, and under-resourced, but the direction is clear. If you are a small vendor, pre-qualify before you bid, because the award cycle will not tolerate a retrofit. The practical takeaway is that program managers who ignore this detail usually see it return as a corrective action.
Vendor attestations are not a one-time document and must be re-signed on each material change, including model fine-tunes, training data additions. New tool integrations also qualify as triggering events requiring immediate re-signature of the attestation. That cadence has caught firms accustomed to a single annual security questionnaire. The practical result is that your DHS account manager needs a model lifecycle calendar that lines up with your release train. Otherwise you will miss a notification obligation and trigger a corrective action. Program managers who treat the attestation as a release artifact rather than a contract artifact come out ahead.
The Signals Intelligence Boundary and FISA 702 Interactions
Shifting to the intelligence boundary, Section 702 of the Foreign Intelligence Surveillance Act is one of the quieter but most consequential interactions with the Homeland Security AI Guidelines. DHS components do not conduct Section 702 collection themselves, but they do receive derived products and run analytics on top of them. Creates a chain-of-custody problem for AI systems that touch those products. The current guidelines require that any analytic touching Section 702 data pass a civil liberties review and that the model artifacts be stored under compartmented controls. These requirements are strict enough that some teams have chosen to run two separate model pipelines rather than mix compartments.
Beyond compartmentation, there is a hard question about query-time use of foreign intelligence data in domestic enforcement workflows. The Fourth Amendment and the Section 702 statutory framework meet at that boundary. The current guidelines do not and cannot resolve the constitutional question, but they do force documentation that preserves the evidentiary record. Teams that bury 702-derived features inside a general-purpose model risk making that record unreadable, which is why the guidance pushes for explicit lineage tracking. That lineage is often the hardest engineering task in a compliant pipeline.
The last point to make here is that the 702 interaction is where DHS, the Department of Justice. The Office of the Director of National Intelligence must coordinate, and the coordination mechanism is still immature. New joint procedures signaled late in 2026 promise a tighter review, but until those procedures stabilize, a program should assume higher scrutiny than the guidelines alone suggest. For vendors, this means a product that touches intelligence data needs a privacy and civil liberties officer in the room during design review, not after. Treating 702 as a late-stage compliance box is a reliable way to miss a required disclosure.
Red Teaming, Evaluation, and Documentation Expected of DHS AI Owners
Red teaming for DHS AI systems is now an expectation, not an option, and the current guidelines treat the red team report as a required artifact for high-impact AI. The expected red team exercise covers prompt injection, jailbreak resistance, bias probing, and in the case of agentic systems, tool misuse. Reports are stored with the Impact Assessment and are reviewed by the AI Oversight Board, with components graded on remediation closure rate. Weak remediation closure is where many pilots stumble, because the red team finds dozens of issues and the program team does not have engineering budget to close them. That shift is quiet but material for anyone preparing a program schedule. Vendors who treat the obligation as a release artifact rather than a one-time filing come out ahead.
Beyond red teaming, routine evaluation suites are expected to run on every release and must include regression tests against civil rights-sensitive scenarios. These suites are a close cousin of the responsible disclosure discipline that private firms have adopted, as described in our analysis of responsible AI governance frameworks. The common failure mode is treating evaluation as a one-shot gate, when it is a continuous process that generates evidence an inspector general can read in six months. Program teams that invest in continuous evaluation also tend to catch model drift before it becomes a reportable incident. The pattern is now consistent across components and should be assumed as the default posture.
The Future of Homeland Security AI Guidelines Through 2027 and Beyond
Looking ahead, the guidelines will probably move from a stack of documents to a tightly woven regime with hard rulemaking. Likely through a mix of amended Federal Acquisition Regulation clauses and new DHS rulemakings in late 2026 and 2027. The near-term pressure comes from Congress, which has signaled impatience with voluntary guidance and is drafting bills that would make Impact Assessments mandatory by statute. If those bills pass, the current voluntary language in the DHS roadmap becomes binding, and the department will need to reclassify a long tail of tools. Vendors who treat current guidance as optional will be caught unprepared.
Beyond the legislative pressure, operational pressure will come from the growth of agentic AI inside DHS, which the current guidelines only partially address. New revisions are expected to codify the agentic AI patterns from CISA, formalize red team requirements for autonomous workflows, and tighten the Section 702 interaction rules. These revisions will shift costs toward model owners and away from pure platform vendors, which could compress margins for firms that sell closed-model APIs without evaluation harnesses. Our coverage of innovative magnetic navigation beyond GPS is a reminder. Resilience planning for AI must sit next to resilience planning for the sensors and infrastructure that feed it. Teams that build this reality into their roadmap avoid the painful retrofit later.
Finally, the civil rights conversation will shape the next version of the guidelines in a way the current one only hints at. Litigation is pressing specific disclosures and audit rights into the public record. We expect the next revision to add an auditable complaint and redress channel, explicit disclosure requirements for high-impact enforcement AI, and sharper lines around biometric expansion. Boards and audit committees that assume today’s language will survive into 2027 will be surprised. The strongest posture is to design toward the stricter outcome now, because retrofitting a program under a new rule is far more expensive than engineering it in.
Chart From AIplusInfo
DHS AI Use Case Inventory by Component
Count of distinct AI use cases registered in the public DHS inventory, with a toggle to a biometric and enforcement subset view.
Source: counts derived from the DHS AI use case inventory and reporting from Fedscoop on DHS AI inventory growth.
Key Insights on the Current Framework
- DHS now reports over 180 AI use cases across components, visible on the official DHS AI page, with law enforcement and biometrics leading the catalog.
- According to Bloomberg’s reporting on Mobile Fortify’s 1.2 billion-image database, DHS is running one of the largest known operational face-matching deployments inside the United States government today.
- The DHS compliance plan for OMB M-25-21 is a 68-page document, published September 2025, that locks the high-impact classification process.
- Enterprise AI gateways are now the preferred DHS procurement pattern, as Kiteworks on DHS gateway strategy argues, reshaping how vendors package their models.
- According to Fedscoop’s inventory reporting, ICE drives DHS AI use case growth, and the component’s catalog grew faster than any other DHS component through 2025 and early 2026.
- CISA’s joint guidance on AI in operational technology introduces specific controls for agent action logging and tool allowlists, standards most legacy OT environments do not yet meet.
- Civil society analysis from the American Immigration Council on invisible gatekeepers finds that the human-in-the-loop test routinely becomes a formality, which drives the current civil rights litigation strategy.
- Legal analysis from Morrison Foerster on the DHS framework estimates most critical infrastructure operators lag 12 to 18 months behind the agentic AI controls now expected.
Reading the insights together, the guidelines are no longer a single roadmap document and should be approached as a connected regime. The roadmap sets priorities, OMB M-25-21 sets minimum practices, and Directive 139-08 binds the department internally. CISA guidance pushes operators of critical infrastructure to adopt controls that match agentic AI reality. Where the regime still falls short is civil rights redress and biometric accountability, both of which are now moving through litigation and congressional hearings. Operators and vendors who read all four pillars together, instead of each in isolation, find a program that is clearer than any single document suggests. The next twelve months will determine whether that regime hardens into enforceable rules or stays as a demanding but voluntary floor.
| Dimension | DHS AI Roadmap | OMB M-25-21 | Directive 139-08 | CISA Critical Infra Guidance |
|---|---|---|---|---|
| Primary audience | DHS components | All federal agencies | DHS internal workforce | Critical infrastructure operators |
| Legal status | Internal strategy | Executive branch policy | Internal directive | Nonbinding guidance |
| Covers vendors | Indirect | Via contract clauses | Explicit flow-down | Implicit through contracts |
| Impact Assessment required | Suggested | Yes for high-impact | Yes for high-impact | Recommended |
| Biometric treatment | Mentioned | High-impact category | Prohibited uses listed | Not primary focus |
| Red teaming expectation | Encouraged | Minimum practice | Required for high-impact | Recommended for OT agents |
| Civil rights review | Named | Named | Signed by privacy officer | Addressed via operator obligations |
| Enforcement vehicle | Program oversight | Agency compliance plan | Administrative action | Insurance, examiners, litigation |
Real-World Examples of DHS AI Deployments Under the Current Guidelines
CBP Mobile Fortify Face-Matching Rollout
Customs and Border Protection deployed Mobile Fortify to officers in the field in late 2024 and expanded the pilot across more than 1,000 devices by early 2026. Reporting from Bloomberg’s investigation of Mobile Fortify documents a 1.2 billion-image reference database and an operational match-time of under five seconds for routine queries. The deployment produced a measurable workflow lift, with CBP reporting faster identity verification during secondary inspection and a lower false-rejection rate than its prior laptop-based tool. The limitation is substantial, since civil rights groups argue the system should be classified high-impact and subjected to a published Impact Assessment that CBP has only partially released. Vendors watching the rollout should treat Mobile Fortify as the template for how a biometric AI tool is now scoped, cleared, and defended inside DHS. The example makes clear that scale and speed do not substitute for documentation, and that the next contested rollout will be judged against the same bar.
ICE Investigative Drafting Assistant
Immigration and Customs Enforcement rolled out an investigative drafting assistant across its Homeland Security Investigations units during 2025. The tool handling an estimated 15 percent of field investigation report drafts by year-end. Fedscoop’s analysis of the DHS AI inventory covering Mobile Fortify and Palantir integrations catalogs the drafting assistant and other ICE tools as middle-tier use cases. The outcome is a reduction in agent paperwork hours per report, estimated at 40 minutes per draft, which translates into thousands of agent-hours redirected each quarter. The limitation is that the drafting assistant was contested by advocacy groups who argued the tier classification understates downstream enforcement impact. The example reveals the gap between middle-tier controls and the oversight expected for anything that touches an enforcement outcome. Vendors should expect similar drafting tools to migrate up the tier ladder in 2026 and 2027.
TSA Automated Identity Verification at the Checkpoint
The Transportation Security Administration deployed automated identity verification at more than 230 airport checkpoints by mid-2026, matching the real-time face capture against the ID presented by the traveler. According to U.S. Travel on biometric upgrades, the system cut average identity verification time by over 30 percent in participating lanes. TSA policy preserves an opt-out option in every participating lane, honored at the podium itself. The system cleared a published privacy impact assessment and sits under the TSA Chief AI Officer’s oversight model. The limitation is that the opt-out option is poorly advertised, and travelers often do not realize the opt-out is available until after capture. The example demonstrates that civil liberties guarantees in the guidelines live or die at the user interface rather than at the policy layer. Future rollouts will need to make the opt-out as visible as the capture itself.
Shop From AIplusInfo
Books That Sharpen Your Reading of the DHS AI Guidelines
These three well-reviewed titles unpack the governance questions sitting under the Homeland Security AI framework and are the clearest primers we recommend for policy and compliance teams working the file.
AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference
Narayanan and Kapoor’s framework for separating durable AI capability from marketing is directly applicable to how DHS components should scope high-impact AI claims.
Buy on AmazonWeapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy
O’Neil’s examination of high-stakes algorithmic systems maps directly to the civil rights questions driving DHS high-impact AI classification debates.
Buy on AmazonThe Alignment Problem: Machine Learning and Human Values
Brian Christian’s accessible synthesis of ML safety, fairness, and oversight is the clearest primer for the civil liberties concerns central to the DHS guidelines debate.
Buy on AmazonAs an Amazon Associate, AIplusInfo earns from qualifying purchases.
Case Studies That Reveal How the Guidelines Hold Up Under Pressure
Case Study: Palantir Integration Across ICE Case Analytics
Immigration and Customs Enforcement had a long-standing problem combining lead generation, case prioritization, and investigation support across multiple systems that did not share a common data model. ICE deployed and expanded its Palantir-based solution through 2024 and 2025 to become the connective layer across investigations, lead triage. Reporting, which Fedscoop’s inventory analysis describes as the single largest component-level AI integration inside DHS. The reported measurable impact includes a double-digit reduction in investigation cycle time for certain case types and the ability to consolidate records that previously lived in dozens of local databases. The controversy is sharp, because advocacy groups argue the platform centralizes sensitive data at a scale the Impact Assessment process did not fully anticipate. The platform’s documentation remains partially redacted, which frustrates outside review and keeps the controversy live.
The guideline tension here is subtle because Palantir itself does not make enforcement decisions, so the component argues the middle-tier classification is correct. Civil rights organizations counter that the composite platform generates lead lists and risk scores that functionally direct enforcement attention, which places it closer to the high-impact category. The case study reveals that composite systems blur tier boundaries, and that the guidelines will likely need a composite-system classification by 2027. Program teams rolling out similar integrations should document composite behavior explicitly, since the alternative is to be reclassified under subpoena. That is the pattern most inspectors general reference during a lookback review.
Case Study: CISA Water Sector Open Source AI Guidance Pilot
Operators in the water and wastewater sector faced a specific problem: how to adopt open source AI models for anomaly detection without exposing operational technology to novel supply chain risk. CISA deployed a solution via joint guidance in mid-2026 targeted at this sector. As covered by Federal News Network on CISA water sector guidance, and ran a pilot with roughly a dozen participating utilities. The pilot provided a shared evaluation harness, a reference model risk inventory, and a shared incident reporting channel. The measurable impact: participating utilities reported faster incident response time and clearer visibility into model behavior, with early metrics suggesting a 25 percent reduction in false positives for anomaly alerts. The pilot’s limitation is coverage, because only a tiny fraction of the sector’s 50,000 systems participated and no mechanism forces broader adoption. Guidance that stays voluntary tends to reach the operators that need it least.
Follow-up conversations with utility CISOs show that even for pilot participants, closing the loop with existing operational technology monitoring remained the hardest step. The pilot’s framework assumes a logging posture that smaller utilities do not have, which is a gap the next guidance revision will need to address. Peer utilities now cite the pilot as the de facto reference for open source AI adoption, which stretches thin guidance into heavier practical weight. The case study illustrates that voluntary guidance can set the market standard, provided a few credible operators adopt it visibly and share the evidence. Programs that invest early in this muscle report fewer surprise findings during their next cycle.
Case Study: USCIS Pre-Screening Model Reclassification
United States Citizenship and Immigration Services operated an asylum pre-screening support tool that initially launched under a middle-tier classification, which generated a short Impact Assessment and limited public documentation. The agency deployed a reclassification solution in 2025 after legal review and civil society pressure, which triggered a fuller Impact Assessment, a civil rights review, and renewed documentation obligations. The documented measurable impact of the USCIS pre-screening reclassification turned out to be large and operationally painful. Filings referenced in Brennan Center’s DHS AI transparency analysis describe a nine-month pause while new assessment work was completed. The agency reported that the heavier evaluation surfaced a bias pattern in the model’s handling of certain country-of-origin cohorts. The direct impact drove remediation through additional training data and threshold adjustments. The limitation is that the reclassification happened late in the lifecycle, which created real operational cost that an earlier classification would have avoided.
Beyond the direct cost, the case study revealed how classification choices propagate downstream, since the pre-screening tool fed into interviews that affected roughly 12 percent of real benefit outcomes. Program managers across DHS watched the reclassification and now treat initial tier choices as consequential rather than administrative. The guidelines explicitly allow for reclassification over time, but the USCIS episode shows that waiting for pressure is the most expensive path. The practical lesson is to classify at the ceiling of plausible downstream use, then step down with evidence, instead of the reverse. Teams that follow this pattern report fewer surprise pauses and more predictable program timelines.
Risks, Ethical Tensions, and Oversight Gaps in the Current Framework
Shifting to risks, the the current guidelines face three serious tensions that will shape the next cycle of revisions. The first tension is civil rights and biometric accountability, where the public inventory and the practical deployment of face-matching systems have outrun the redress channel. Advocacy groups argue the current process does not give an individual a realistic way to challenge an AI-affected enforcement decision, and litigation is pressing that gap into the record. The expected outcome is a formal, independent redress mechanism inside DHS, with a public dashboard tracking complaints, outcomes, and remediation. The lesson for 2026 and 2027 is to design for scrutiny from day one, not from the first audit.
The second tension is the pace of agentic AI adoption, which the guidelines partially address but which still outruns the department’s ability to red team and monitor. New tools are shipped faster than Impact Assessments can be refreshed, and components face pressure to use agentic patterns before CISA’s agentic AI document can be fully internalized. Analysts tracking autonomous AI oversight frameworks argue that agentic AI needs a dedicated classification tier because its risks are structurally different from traditional model risks. The current single-tier approach may need to split into separate agentic and non-agentic rails to remain workable. Market analysts expect this trend to accelerate as rulemaking tightens the loop.
The third tension is institutional capacity across the AI governance offices. The AI Oversight Board, component Chief AI Officers, and privacy and civil rights offices are all under-resourced relative to use case inventory growth. That gap is particularly acute for independent red teaming, where the pool of cleared, qualified reviewers has not grown as fast as the pool of systems requiring review. The result is a quiet triage, where some high-impact systems get first-class red team coverage and others get a lighter review that risks missing structural issues. Closing the capacity gap is a budgetary and workforce question. The next revision of the guidelines may need to pair stronger expectations with a stronger resource commitment, otherwise the policies will exist on paper but not in practice. The guidance is still evolving, so document any choice carefully so your next revision can trace it.
Frequently Asked Questions on Homeland Security AI Guidelines
The Homeland Security AI Guidelines are the layered DHS, OMB, and CISA documents that govern how federal and critical infrastructure AI systems are built, assessed, documented, and operated. They combine the DHS AI Roadmap, OMB M-25-21 compliance requirements, Directive 139-08, and CISA sector guidance. Together they set risk tiers, Impact Assessment duties, and vendor obligations for every DHS component and contractor.
Compliance obligations fall on every DHS component, their contractors, and operators across the sixteen critical infrastructure sectors defined by CISA. Any AI system that influences safety, rights, or security decisions sits within scope. Vendors that sell models or AI-enabled platforms into DHS inherit flow-down obligations through standard contract clauses now in use.
High-impact AI is any system whose decisions materially affect rights, safety, access to benefits, or law enforcement outcomes. The classification triggers mandatory Impact Assessments, Chief AI Officer sign-off, and quarterly AI Oversight Board review. Facial recognition, benefit adjudication, and enforcement-adjacent tools typically fall inside the high-impact band.
Directive 139-08 is the DHS internal directive that formalizes the Chief AI Officer role, assigns an AI Oversight Board, defines prohibited AI uses, and ties AI governance to acquisition clauses. The directive officially took effect across all DHS components in January 2025 after internal review. The directive is the single most binding internal document across the policy stack.
OMB M-25-21 requires every federal agency to publish an AI compliance plan, maintain a public use case inventory, and define a high-impact classification process. The memo mandates minimum practices for pre-deployment testing of covered AI systems. DHS released its 68-page compliance plan in September 2025, with formal waiver procedures and tighter vendor flow-down requirements.
Mobile Fortify is a handheld face-matching application used by DHS immigration officers that reads against a 1.2 billion-image database. Critics argue it is high-impact AI that should trigger a published Impact Assessment and robust civil rights review. The tool is at the center of ongoing litigation and advocacy challenges in 2026.
CISA sets AI security guidance for critical infrastructure operators, with safety and security guidelines from 2024, agentic AI guidance in 2026, and OT-specific joint guides issued throughout 2025 and 2026. The guidance is nonbinding in law but is cited by insurers, examiners, and plaintiffs. Many operators therefore treat the guidance as a de facto rule.
An Impact Assessment is a structured document that captures the problem, the model, the training data, the evaluation results, and a civil rights review signed by the component privacy officer. It is required for high-impact AI and re-run on material change. The Impact Assessment is reviewed by the AI Oversight Board.
Agencies had to publish compliance plans by late September 2025 under OMB M-25-21. Component-level use case inventories update quarterly, with the next AI Oversight Board full review window landing in early 2026. Expected rulemakings in late 2026 and 2027 may convert current voluntary guidance into binding regulations.
Yes, vendors can sell AI products to DHS without a Chief AI Officer, but the contract clauses still require named owners for data governance, model updates, and evaluation harness upkeep. Mid-size vendors typically assign these duties to engineering and legal leads. Pre-qualifying through the new DHS evaluation pathway reduces award cycle risk.
CISA’s 2026 agentic AI guidance pushes operators to inventory tool permissions, enforce allowlists, log every agent action, and require human-in-the-loop confirmation for destructive actions. The guidance addresses prompt injection and confused deputy risks explicitly. Operators are expected to add tamper-evident logging that most legacy systems do not provide today.
The guidelines call for a complaint and redress channel where individuals can challenge AI-affected enforcement or benefit decisions. The channel is new and uptake has been uneven, with civil society groups arguing it needs stronger independence and visibility. Litigation is pushing DHS toward a formal, dashboard-tracked redress mechanism.
Yes, DHS components that receive Section 702-derived products must apply civil liberties review, compartmented storage, and explicit lineage tracking for any model that touches those products. The guidelines do not and cannot resolve the underlying constitutional question. They do require evidentiary documentation that preserves the record for later review.
Vendor attestations are tied to material changes, including model fine-tunes, training data additions, and new tool integrations. Firms shipping weekly or monthly refreshes must notify DHS before each release ships, with evidence that the evaluation suite still passes. Treating attestations as release artifacts rather than annual security questionnaires avoids corrective actions.
The DHS AI page hosts the roadmap, the use case inventory, and links to Directive 139-08. OMB M-25-21 and the DHS compliance plan are published on the DHS and White House websites. CISA guidance lives on cisa.gov, and sector-specific guidance appears in the agency’s news and resources section.